| | |
| | | #include <stdio.h> |
| | | #include <time.h> |
| | | |
| | | void swap_binary(convolutional_layer *l) |
| | | { |
| | | float *swap = l->filters; |
| | | l->filters = l->binary_filters; |
| | | l->binary_filters = swap; |
| | | |
| | | #ifdef GPU |
| | | swap = l->filters_gpu; |
| | | l->filters_gpu = l->binary_filters_gpu; |
| | | l->binary_filters_gpu = swap; |
| | | #endif |
| | | } |
| | | |
| | | void binarize_filters2(float *filters, int n, int size, char *binary, float *scales) |
| | | { |
| | | int i, k, f; |
| | | for(f = 0; f < n; ++f){ |
| | | float mean = 0; |
| | | for(i = 0; i < size; ++i){ |
| | | mean += fabs(filters[f*size + i]); |
| | | } |
| | | mean = mean / size; |
| | | scales[f] = mean; |
| | | for(i = 0; i < size/8; ++i){ |
| | | binary[f*size + i] = (filters[f*size + i] > 0) ? 1 : 0; |
| | | for(k = 0; k < 8; ++k){ |
| | | } |
| | | } |
| | | } |
| | | } |
| | | |
| | | void binarize_filters(float *filters, int n, int size, float *binary) |
| | | { |
| | | int i, f; |
| | | for(f = 0; f < n; ++f){ |
| | | float mean = 0; |
| | | for(i = 0; i < size; ++i){ |
| | | mean += fabs(filters[f*size + i]); |
| | | } |
| | | mean = mean / size; |
| | | for(i = 0; i < size; ++i){ |
| | | binary[f*size + i] = (filters[f*size + i] > 0) ? mean : -mean; |
| | | } |
| | | } |
| | | } |
| | | |
| | | int convolutional_out_height(convolutional_layer l) |
| | | { |
| | | int h = l.h; |
| | |
| | | return float_to_image(w,h,c,l.delta); |
| | | } |
| | | |
| | | convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize) |
| | | void backward_scale_cpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates) |
| | | { |
| | | int i,b,f; |
| | | for(f = 0; f < n; ++f){ |
| | | float sum = 0; |
| | | for(b = 0; b < batch; ++b){ |
| | | for(i = 0; i < size; ++i){ |
| | | int index = i + size*(f + n*b); |
| | | sum += delta[index] * x_norm[index]; |
| | | } |
| | | } |
| | | scale_updates[f] += sum; |
| | | } |
| | | } |
| | | |
| | | void mean_delta_cpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta) |
| | | { |
| | | |
| | | int i,j,k; |
| | | for(i = 0; i < filters; ++i){ |
| | | mean_delta[i] = 0; |
| | | for (j = 0; j < batch; ++j) { |
| | | for (k = 0; k < spatial; ++k) { |
| | | int index = j*filters*spatial + i*spatial + k; |
| | | mean_delta[i] += delta[index]; |
| | | } |
| | | } |
| | | mean_delta[i] *= (-1./sqrt(variance[i] + .00001f)); |
| | | } |
| | | } |
| | | void variance_delta_cpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta) |
| | | { |
| | | |
| | | int i,j,k; |
| | | for(i = 0; i < filters; ++i){ |
| | | variance_delta[i] = 0; |
| | | for(j = 0; j < batch; ++j){ |
| | | for(k = 0; k < spatial; ++k){ |
| | | int index = j*filters*spatial + i*spatial + k; |
| | | variance_delta[i] += delta[index]*(x[index] - mean[i]); |
| | | } |
| | | } |
| | | variance_delta[i] *= -.5 * pow(variance[i] + .00001f, (float)(-3./2.)); |
| | | } |
| | | } |
| | | void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta) |
| | | { |
| | | int f, j, k; |
| | | for(j = 0; j < batch; ++j){ |
| | | for(f = 0; f < filters; ++f){ |
| | | for(k = 0; k < spatial; ++k){ |
| | | int index = j*filters*spatial + f*spatial + k; |
| | | delta[index] = delta[index] * 1./(sqrt(variance[f]) + .00001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch); |
| | | } |
| | | } |
| | | } |
| | | } |
| | | |
| | | convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize, int binary) |
| | | { |
| | | int i; |
| | | convolutional_layer l = {0}; |
| | |
| | | l.w = w; |
| | | l.c = c; |
| | | l.n = n; |
| | | l.binary = binary; |
| | | l.batch = batch; |
| | | l.stride = stride; |
| | | l.size = size; |
| | |
| | | l.output = calloc(l.batch*out_h * out_w * n, sizeof(float)); |
| | | l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float)); |
| | | |
| | | if(binary){ |
| | | l.binary_filters = calloc(c*n*size*size, sizeof(float)); |
| | | l.cfilters = calloc(c*n*size*size, sizeof(char)); |
| | | l.scales = calloc(n, sizeof(float)); |
| | | } |
| | | |
| | | if(batch_normalize){ |
| | | l.scales = calloc(n, sizeof(float)); |
| | | l.scale_updates = calloc(n, sizeof(float)); |
| | |
| | | l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n); |
| | | l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n); |
| | | |
| | | if(binary){ |
| | | l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size); |
| | | } |
| | | |
| | | if(batch_normalize){ |
| | | l.mean_gpu = cuda_make_array(l.mean, n); |
| | | l.variance_gpu = cuda_make_array(l.variance, n); |
| | |
| | | |
| | | void test_convolutional_layer() |
| | | { |
| | | convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1); |
| | | convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 0); |
| | | l.batch_normalize = 1; |
| | | float data[] = {1,1,1,1,1, |
| | | 1,1,1,1,1, |
| | |
| | | } |
| | | } |
| | | |
| | | void forward_convolutional_layer(const convolutional_layer l, network_state state) |
| | | void forward_convolutional_layer(convolutional_layer l, network_state state) |
| | | { |
| | | int out_h = convolutional_out_height(l); |
| | | int out_w = convolutional_out_width(l); |
| | | int i; |
| | | |
| | | fill_cpu(l.outputs*l.batch, 0, l.output, 1); |
| | | /* |
| | | if(l.binary){ |
| | | binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters); |
| | | binarize_filters2(l.filters, l.n, l.c*l.size*l.size, l.cfilters, l.scales); |
| | | swap_binary(&l); |
| | | } |
| | | */ |
| | | |
| | | if(l.binary){ |
| | | int m = l.n; |
| | | int k = l.size*l.size*l.c; |
| | | int n = out_h*out_w; |
| | | |
| | | char *a = l.cfilters; |
| | | float *b = l.col_image; |
| | | float *c = l.output; |
| | | |
| | | for(i = 0; i < l.batch; ++i){ |
| | | im2col_cpu(state.input, l.c, l.h, l.w, |
| | | l.size, l.stride, l.pad, b); |
| | | gemm_bin(m,n,k,1,a,k,b,n,c,n); |
| | | c += n*m; |
| | | state.input += l.c*l.h*l.w; |
| | | } |
| | | scale_bias(l.output, l.scales, l.batch, l.n, out_h*out_w); |
| | | add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w); |
| | | activate_array(l.output, m*n*l.batch, l.activation); |
| | | return; |
| | | } |
| | | |
| | | int m = l.n; |
| | | int k = l.size*l.size*l.c; |